A Transformer-Based Approach for 3D Human Pose Estimation in Rehabilitation Exercise Movements
Naichen Kang, Gang Chen, Cong Zhang, Yubo Xue · 2024
The proportion of elderly patients with chronic diseases, represented by stroke, continues to rise. More and more elderly patients in the rehabilitation stage are facing serious challenges such as long rehabilitation periods, difficulty perceiving the effects of the rehabilitation stage, and limited mobility leading to difficulties in offline visits. These issues undoubtedly create significant psychological pressure and economic burden on patients. In response to this situation, this paper proposes a multistage 3D human pose estimation method based on a monocular camera. The model is specifically designed to address common environmental occlusion and inter-limb occlusion issues, aiming to improve the measurement accuracy and evaluation effectiveness of the model. The network structure is based on a combination model of HRNet+OCNet to achieve high-precision prediction of 3D keypoints. By incorporating a correction module based on PAF, the coordinates of target keypoints under “traversal occlusion” and “ full occlusion of end joint points “ scenes are corrected, thereby obtaining a 3D human pose sequence closer to reality.